First steps / The idea

What a surrogate is

A fast, learned map from a design to its physics, from a few numbers to the full field.

A surrogate model is a fast, learned approximation of an expensive simulation or experiment. Trained on solver results, it predicts quantities of interest, or full physical fields such as pressure, velocity and temperature, for new designs in milliseconds instead of hours. Surrogate models are used in engineering design, optimisation, uncertainty quantification and digital twins.

Common questions

What is a surrogate model?

A surrogate model is a fast, learned approximation of an expensive simulation or experiment. Trained on solver results, it predicts quantities of interest, or full physical fields such as pressure, velocity and temperature, for new designs in milliseconds instead of hours.

What types of surrogate models are there?

Scalar surrogates (response surfaces, kriging or Gaussian processes, reduced-order models) predict a few numbers. Full-field neural surrogates predict whole fields: on grids (U-Net, Fourier neural operators), on graphs (MeshGraphNets) and on point clouds (neural fields, Transolver, GINO), increasingly with latent transformers.

What is the difference between a surrogate model and a simulation?

A simulation solves the governing equations for each new design, which can take hours. A surrogate learns from many solved designs and then answers instantly, with an error that must be validated and that grows outside the training range.

What are surrogate models used for?

Design-space exploration, optimisation, uncertainty quantification, sensitivity analysis and real-time digital twins, in aerospace, automotive, energy, electronics and manufacturing.

How accurate are surrogate models?

Inside the range of their training data they typically reproduce the simulation within a few percent; outside it they extrapolate without warning. Accuracy is measured on held-out designs and reported with the training range.

Scalar surrogates predict a few numbers such as lift, drag or peak temperature; full-field surrogates predict the whole field on the geometry, using grids, graphs or point clouds.

Related

• Where it fits: Tests validate simulations, simulations train the surrogate, the surrogate picks what to simulate next.
• Representations: Voxel grids, graphs or point clouds: how geometry and fields become numbers.
• Model families: Neural fields, U-Nets, graph networks and transformers, and what each is good for.